- 18 Apr, 2019 36 commits
-
-
Mika Westerberg authored
Titan Ridge flow to start the firmware is the same as Alpine Ridge so we can do the same on Titan Ridge based Apple systems. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
While tb_dump_hop() prints out necessary information it is in format that is quite hard to read from the logs especially when one needs to follow the path to see that the setup is correct. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
Now that the driver can handle every possible tunnel types there is no point to log everything as info level so turn these to happen at debug level instead. While at it remove duplicated tunnel activation log message (tb_tunnel_activate() calls tb_tunnel_restart() which print the same message) and add one missing '\n' termination. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
The printing macros do not modify the passed object so make them const. While there make tb_route() to take const parameter as well. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
Two domains (hosts) can be connected through a Thunderbolt cable and in that case they can start software services such as networking over the high-speed DMA paths. Now that we have all the basic building blocks in place to create DMA tunnels over the Thunderbolt fabric we can add this support to the software connection manager as well. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
In order to detect possible connections to other domains we need to be able to find out why tb_switch_alloc() fails so make it return ERR_PTR() instead. This allows the caller to differentiate between errors such as -ENOMEM which comes from the kernel and for instance -EIO which comes from the hardware when trying to access the possible switch. Convert all the current call sites to handle this properly. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
In addition to PCIe and Display Port tunnels it is also possible to create tunnels that forward DMA traffic from the host interface adapter (NHI) to a NULL port that is connected to another domain through a Thunderbolt cable. These tunnels can be used to carry software messages such as networking packets. To support this we introduce another tunnel type (TB_TUNNEL_DMA) that supports paths from NHI to NULL port and back. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
Currently ICM has been handling XDomain UUID exchange so there was no need to have it in the driver yet. However, since now we are going to add the same capabilities to the software connection manager it needs to be handled properly. For this reason modify the driver XDomain protocol handling so that if the remote domain UUID is not filled in the core will query it first and only then start the normal property exchange flow. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
We run all XDomain requests during discovery in tb->wq and since it only runs one work at the time it means that sending back reply to the other domain may be delayed too much depending whether there is an active XDomain discovery request running. To make sure we can send reply to the other domain as soon as possible run tb_xdp_handle_request() in system workqueue instead. Since the device can be hot-removed in the middle we need to make sure the domain structure is still around when the function is run so increase reference count before we schedule the reply work. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
Now that we have capability to discover existing tunnels during driver load there is no point tearing down tunnels when the driver gets unloaded. Instead we can just leave them running. If user disconnects devices while there is no Thunderbolt driver loaded, tunneled protocol hotplug happens and is handled by the corresponding driver (pciehp in case of PCIe tunnel, GFX driver in case of DP tunnel). Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
Display Port tunnels are somewhat more complex than PCIe tunnels as it requires 3 tunnels (AUX Rx/Tx and Video). In addition we are not supposed to create the tunnels immediately when a DP OUT is enumerated. Instead we need to wait until we get hotplug event to that adapter port or check if the port has HPD set before tunnels can be established. This adds Display Port tunneling support to the software connection manager. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
NFC (non flow control) credits is actually 20-bit field so update tb_port_add_nfc_credits() to handle this properly. This allows us to set NFC credits for Display Port path in subsequent patches. Also make sure the function does not update the hardware if the underlying switch is already unplugged. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
We will be needing these routines to find Display Port adapters as well so modify them to take port type as the second parameter. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
The only way to expand Thunderbolt topology is through the NULL adapter ports (typically ports 1, 2, 3 and 4). There is no point handling Thunderbolt hotplug events on any other port. Add a helper function (tb_port_is_null()) that can be used to determine if the port is NULL port, and use it in software connection manager code when hotplug event is handled. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
Currently the software connection manager (tb.c) has only supported creating a single PCIe tunnel, no PCIe device daisy chaining has been supported so far. This updates the software connection manager so that it now can create PCIe tunnels for full chain of six devices. Because PCIe allows DMA and opens possibility for DMA attacks we change security level to "user" meaning that PCIe tunneling requires that the userspace authorizes the devices first. This makes it possible to block PCIe tunneling completely while still allowing other types of tunnels to be automatically created. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
In Apple Macs the boot firmware (EFI) connects all devices automatically when the system is started, before it hands over to the OS. Instead of ignoring we discover all those PCIe tunnels and record them using our internal structures, just like we do when a device is connected after the OS is already up. By doing this we can properly tear down tunnels when devices are disconnected. Also this allows us to resume the existing tunnels after system suspend/resume cycle. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
State of the connected devices and tunnel configuration is not known during resume. For example some paths may not be complete anymore if the user has unplugged the related devices. So instead of marking all paths as inactive we go ahead and deactivate them explicitly before we restart them. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
Now that we can allocate hop IDs per port on a path, we can take advantage of this and create tunnels covering longer paths than just between two adjacent switches. PCIe actually does not need this as it is typically a daisy chain between two adjacent switches but this way we do not need to hard-code creation of the tunnel. While there add name to struct tb_path to make debugging easier, and update kernel-doc comments. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
We need to be able to walk from one port to another when we are creating paths where there are multiple switches between two ports. For this reason introduce a new function tb_next_port_on_path(). Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com> Reviewed-by: Lukas Wunner <lukas@wunner.de>
-
Mika Westerberg authored
Currently the driver only assigns remote port for the primary port if in case of dual link. This makes things such as walking from one port to another more complex than necessary because the code needs to change from secondary to primary port if the path that is established is created using secondary links. In order to always assign both remote pointers we need to prevent the scanning code from following the secondary link. Failing to do that might cause problems as the same switch may be enumerated twice (or removed in case of unplug). Handle that properly by introducing a new function tb_port_has_remote() that returns true only for the primary port. We also update tb_is_upstream_port() to support both dual link ports, make it take const port pointer and move it below tb_upstream_port() to keep similar functions close. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
Each port has a separate path configuration space that is used for finding the next hop (switch) in the path. HopID is an index to this configuration space. HopIDs 0 - 7 are reserved by the protocol. In order to get next available HopID for each direction we provide two pairs of helper functions that can be used to allocate and release HopIDs for a given port. While there remove obsolete TODO comment. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
To be able to tunnel non-PCIe traffic, separate tunnel functionality into generic and PCIe specific parts. Rename struct tb_pci_tunnel to tb_tunnel, and make it hold an array of paths instead of just two. Update all the tunneling functions to take this structure as parameter. We also move tb_pci_port_active() to switch.c (and rename it) where we will be keeping all port and switch related functions. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
In order to tunnel non-PCIe traffic as well rename tunnel_pci.[ch] to tunnel.[ch] to reflect this fact. No functional changes. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
The adapter specific capability either is there or not if the port does not hold an adapter. Instead of always finding it on-demand we read the offset just once when the port is initialized. While there we update the struct port documentation to follow kernel-doc format. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
We need to wait until all buffers have been drained before the path can be considered disabled. Do this for every hop in a path. This adds another bit field to struct tb_regs_hop even if we are trying to get rid of them but we can clean them up another day. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
Thunderbolt 2 devices and beyond link controller needs to be notified when a switch is going to be suspended by setting bit 31 in LC_SX_CTRL register. Add this functionality to the software connection manager. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
Thunderbolt 2 devices and beyond need to have additional bits set in link controller specific registers. This includes two bits in LC_SX_CTRL that tell the link controller which lane is connected and whether it is upstream facing or not. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
We will be adding more link controller functionality in subsequent patches and it does not make sense to keep all that in switch.c, so separate LC functionality into its own file. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
Light Ridge has an issue where reading the next capability pointer location in port config space the read data is not cleared. It is fine to read capabilities each after another so only thing we need to do is to make sure we issue dummy read after tb_port_find_cap() is finished to avoid the issue in next read. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
Light Ridge and Eagle Ridge both need to have TMU access enabled before port space can be fully accessed so make sure it happens on those. This allows us to get rid of the offset quirk in tb_port_find_cap(). Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
Maximum depth in Thunderbolt topology is 6 so make sure it is not possible to allocate switches that exceed the depth limit. While at it update tb_switch_alloc() to use upper/lower_32_bits() following tb_switch_alloc_safe_mode(). Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
switch_lock was introduced because it allowed serialization of device authorization requests from userspace without need to take the big domain lock (tb->lock). This was fine because device authorization with ICM is just one command that is sent to the firmware. Now that we start to handle all tunneling in the driver switch_lock is not enough because we need to walk over the topology to establish paths. For this reason drop switch_lock from the driver completely in favour of big domain lock. There is one complication, though. If userspace is waiting for the lock in tb_switch_set_authorized(), it keeps the device_del() from removing the sysfs attribute because it waits for active users to release the attribute first which leads into following splat: INFO: task kworker/u8:3:73 blocked for more than 61 seconds. Tainted: G W 5.1.0-rc1+ #244 "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. kworker/u8:3 D12976 73 2 0x80000000 Workqueue: thunderbolt0 tb_handle_hotplug [thunderbolt] Call Trace: ? __schedule+0x2e5/0x740 ? _raw_spin_lock_irqsave+0x12/0x40 ? prepare_to_wait_event+0xc5/0x160 schedule+0x2d/0x80 __kernfs_remove.part.17+0x183/0x1f0 ? finish_wait+0x80/0x80 kernfs_remove_by_name_ns+0x4a/0x90 remove_files.isra.1+0x2b/0x60 sysfs_remove_group+0x38/0x80 sysfs_remove_groups+0x24/0x40 device_remove_attrs+0x3d/0x70 device_del+0x14c/0x360 device_unregister+0x15/0x50 tb_switch_remove+0x9e/0x1d0 [thunderbolt] tb_handle_hotplug+0x119/0x5a0 [thunderbolt] ? process_one_work+0x1b7/0x420 process_one_work+0x1b7/0x420 worker_thread+0x37/0x380 ? _raw_spin_unlock_irqrestore+0xf/0x30 ? process_one_work+0x420/0x420 kthread+0x118/0x130 ? kthread_create_on_node+0x60/0x60 ret_from_fork+0x35/0x40 We deal this by following what network stack did for some of their attributes and use mutex_trylock() with restart_syscall(). This makes userspace release the attribute allowing sysfs attribute removal to progress before the write is restarted and eventually fail when the attribute is removed. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
If switch is already disconnected there is no point sending it commands and waiting for timeout. Instead in that case return error immediately. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
tb_switch_find_by_route() does the same already so use it instead and remove duplicated get_switch_at_route(). Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com> Reviewed-by: Lukas Wunner <lukas@wunner.de>
-
Mika Westerberg authored
This field is not used anywhere so remove it. Reported-by: Lukas Wunner <lukas@wunner.de> Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Mika Westerberg authored
The XDomain protocol messages may start as soon as Thunderbolt control channel is started. This means that if the other host starts sending ThunderboltIP packets early enough they will be passed to the network driver which then gets confused because its resume hook is not called yet. Fix this by unregistering the ThunderboltIP protocol handler when suspending and registering it back on resume. Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com> Acked-by: David S. Miller <davem@davemloft.net>
-
- 28 Mar, 2019 2 commits
-
-
Aditya Pakki authored
uuid in add_switch is allocted via kmemdup which can fail. The patch logs the error and cleans up the allocated memory for switch. Signed-off-by: Aditya Pakki <pakki001@umn.edu> Reviewed-by: Mukesh Ojha <mojha@codeaurora.org> Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Kangjie Lu authored
No check is enforced for the return value of kzalloc, which may lead to NULL-pointer dereference. The patch fixes this issue. Signed-off-by: Kangjie Lu <kjlu@umn.edu> Reviewed-by: Mukesh Ojha <mojha@codeaurora.org> Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
- 22 Mar, 2019 2 commits
-
-
Aditya Pakki authored
kmemdup can fail and return a NULL pointer. The patch modifies the signature of tb_xdp_schedule_request and passes the failure error upstream. Signed-off-by: Aditya Pakki <pakki001@umn.edu> Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-
Aditya Pakki authored
In enumerate_services, ida_simple_get on failure can return an error and leaks memory. The patch ensures that the dev_set_name is set on non failure cases, and releases memory during failure. Signed-off-by: Aditya Pakki <pakki001@umn.edu> Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
-